Supercomputing in Transition
Next week the HPC digerati will descend upon the International Supercomputing Conference (ISC) in Hamburg, Germany. As the European counterbalance to the larger Supercomputing Conference (SC) in the US, ISC strives for a more international flavor and a more intimate vibe — although with 2,300-plus attendees and 152 exhibitors registered this year, it’s definitely starting to feel like a smaller version of SC.
Occurring on the cusp of the spring-summer transition, ISC manages to fill a seasonal void for HPC vendors looking to introduce products or talk up new technologies and strategies in the middle of the year. While there will be a flurry of vendor announcements this year, I think a lot of the talk is going to be around technology.
HPC is in one of those fundamental transitions right now as it shifts from monolithic CPU-based systems to heterogeneous platforms. That’s changing the hardware and software fundamentals of high performance computing, not to mention the vendor and geographic supercomputing landscape.
Terms that were alien to HPC a few short years ago — GPU computing, CUDA, OpenCL, APUs, manycore processors, Chinese supercomputers — are now part of the vernacular. With chip giants like Intel, AMD, and NVIDIA all offering (or soon offering) accelerator products for HPC, and software support racing to catch up, we are seeing an architectural shift as profound as the 1990s-era transition from proprietary vector and MPP supercomputers to commodity clusters.
It’s now pretty much an accepted fact that the petascale age will be chock full of heterogeneous computer systems. It’s fairly safe to say that most, if not all, architectures for exascale systems, will have a significant heterogeneous component to them — most likely with on-chip floating point accelerators. Whether these are integrated GPUs, Intel’s Many Integrated Core (MIC) processors, or some other variation of a fat core-thin core design, remains to be seen.
As you might imagine, there will be a number of sessions at ISC on this topic set including a panel that tackles the subject head-on: Heterogeneous Systems & Their Challenges to HPC Systems, which gets under way on Monday afternoon.
Berkeley Lab’s John Shalf hosts the panel, which includes Cray CTO Steve Scott, Intel researcher Pradeep Dubey, the University of Heidelberg’s Rainer Spurzem, and Kai Lu of China’s National University of Defense Technology (NUDT).
With the Cray unveiling of their new XK6 GPU super fresh in his mind, Scott is liable to devote some attention to the joys and heartache of CPU-GPU heterogeneity and the importance of offering a portable, productive, high-level software environment for this environment. Cray is pushing for OpenMP accelerator extensions to fill the void here, so I’d expect that be part of his spiel.
Intel’s Pradeep Dubey was one of the authors of a 2010 paper that aimed to debunk the 100X to 1000X performance improvements claimed for GPUs (compared to CPUs). Given that Dubey is also familiar with Intel’s MIC architecture, he’s likely to draw some distinction between the two approaches, especially in regard to ease of programming.
Rainer Spurzem employs GPUs to accelerate astrophysics applications, so he’ll offer the perspective of heterogeneous computing user. Astrophysics, which until recently was an observational science, now increasingly uses compute-heavy simulations and signal processing as a common tool. GPUs are well-suited to such work, although some researchers have gone the FPGA route. Spurzem could offer some perspective on how best to exploit these accelerators.
Expect NUDT’s Kai Lu to talk up China’s push into heterogenous computing via its embrace of GPU-boosted supercomputers. NUDT is the developer of the reigning TOP500 champ, Tianhe-1A, which clocks in at over 2.5 Linpack petaflops. Although there is plenty of skepticism to go around about the practical utility of such massive GPU-accelerated machines, the Tianhe super recently claimed 1.87 petaflop of performance on a real-world molecular simulation application.
Another session along the same lines is Tuesday’s GPU debate between HPC icon Thomas Sterling and NVIDIA’s David Kirk. Kirk, obviously, will be the advocate for GPU computing, with Sterling there to offer the rebuttal and some historical perspective.
Heterogeneous computing is likely to permeate a number of other ISC presentations, including Monday’s session on Transpeta Flops Initiatives and Wednesday’s session on Many-Core Computing. I can also promise you that there will be breaking vendor news around this topic at ISC, so be sure to catch our special coverage of ISC starting next week.